Quick answer
Input tokens are new text sent to the model. Output tokens are what it writes back. Cache write tokens store a prompt prefix for reuse, and cache read tokens reuse it at a much lower price. In agent tools like Claude Code, cache reads are usually most of your total.
The four token types
- Input: new text the model reads, such as your message or a file it opened.
- Output: text the model writes, including code and tool calls. Usually the most expensive per token.
- Cache write: input stored so later turns can reuse it. Priced a little above normal input on Anthropic models.
- Cache read: stored input reused on a later turn. Priced at a fraction of normal input.
Why your total looks huge
An agent resends the whole conversation on every turn. With caching, most of that resend becomes cache reads. A session can show millions of tokens while the new input and output are a small slice.
So a big token number is not a big bill by itself. Look at the split before you panic.
See your split
UsageAtlas shows input, output, and cache tokens for Claude Code, Codex, OpenCode, and other local sources, then estimates cost per model. Use it to check whether a heavy day came from long context or from a lot of output. More on cost in the AI coding cost tracker guide.
Questions, answered
Do cache read tokens count toward Claude Code limits?
Anthropic does not publish a token formula for plan limits. Watch the limit percentage itself rather than converting tokens to a quota.
Why are cache reads so cheap?
The provider already processed that text, so reusing it costs far less compute than reading it fresh.
Can I reduce cache tokens?
Shorter sessions and clearing context between unrelated tasks shrink what each turn resends.